rsh-review

Reviewer

Review code changes with a bug-finding mindset. Prioritize correctness, regressions, and missing coverage over style nits. Keep feedback specific, actionable, and grounded in the actual diff.

Use This Skill For

• Reviewing a diff, patch, PR, or local code changes
• Looking for bugs, regressions, missing tests, or risky behavior changes
• Producing review findings rather than implementation work

Do Not Use This Skill For

• Writing the implementation itself
• Primary docs-writing tasks
• Design work before code exists

Review Priorities

1. Correctness and behavior changes
2. Regressions in CLI, plugin, auth, spec-loading, output, or config flows
3. Missing or weak tests, especially for edge cases and failure paths
4. Error handling, concurrency, subprocess lifecycle, and cleanup
5. User-facing documentation and design-doc impact when behavior changes
6. Performance, security, and maintainability where they materially affect the change

How To Review

1. Start with git status --short, git diff --stat, and git diff --name-only to map the review surface.
2. Read the diff before surrounding code. Identify the user-visible, protocol-visible, or compatibility impact.
3. Read nearby implementation and tests for changed paths. Compare new behavior with existing conventions.
4. Challenge assumptions around error paths, cancellation, timeouts, cleanup, concurrency, config precedence, and backward compatibility.
5. Check whether tests cover intended behavior and important failure modes.
6. Check whether site/ docs or docs/design/ should change.
7. Keep the review scoped to the current PR unless the user asks for a broader audit. Distinguish required fixes from optional hardening or follow-on cleanup.
8. Report findings first, ordered by severity. Keep summaries brief.

Output Expectations

• Lead with concrete findings, not a general summary.
• Prefer sections in this order: Findings, Open questions / assumptions, Residual risk.
• For each finding, include severity, file/line, problem, impact, and the smallest credible fix.
• Explain the impact: what breaks, leaks, regresses, or becomes confusing.
• If no findings are present, say so explicitly and call out any residual risk or untested areas.
• Do not pad the review with praise or low-value nits unless the user asks for them.
• Do not report speculative issues unless there is a plausible failure mode in the changed code.
• Do not turn a scoped PR review into general codebase cleanup. If you notice unrelated risks, list them as out-of-scope follow-ups only when they are important.
• Treat missing coverage as a test gap unless the diff shows a confirmed bug.

Severity Guide

• P0: breakage, data loss, or a serious security issue
• P1: likely regression or incorrect behavior
• P2: test gap, maintainability issue, or edge-case risk with plausible impact
• P3: minor issue, clarity problem, or polish item

Restish-Specific Watchlist

These are high-value repo-specific checks, not an exhaustive checklist. Use them when relevant; do not force them into unrelated reviews. For the full catalog of historical bug classes found in past releases, see ../rsh-release-qa/references/findings-mining.md. When a changed subsystem has a design doc (docs/design/README.md), check the diff against its stated invariants.

Path-specific cues

• internal/cli/: flags, exit codes, stdin/stdout/stderr, config precedence, generated commands
• internal/request/: content negotiation, auth headers, redirects, pagination, retries, streaming, cancellation
• internal/output/: formatter drift, regression fixtures, terminal width behavior, stable ordering
• internal/config/: defaults, file permissions, migrations, backward compatibility
• internal/plugin/ and cmd/restish-*: plugin protocol, subprocess lifecycle, wire compatibility
• cmd/restish and site/: user-facing CLI behavior, examples, docs impact

Subprocess lifecycle

Every exec.Cmd that has been Start()ed must be Wait()ed, or the process becomes a zombie. Error paths usually need cleanup that closes pipes, kills the process if needed, and then waits for exit.

Timeouts around blocking pipe reads

If a goroutine reads from a subprocess pipe and another goroutine times out, the reader can leak forever unless the read is made interruptible. Closing the relevant io.ReadCloser and draining result channels is the usual fix.

Blocking reads from stdin or other io.Readers

A goroutine blocked on Read often cannot be stopped by closing some other resource. Prefer patterns where the blocking read is isolated and the coordinating goroutine can exit cleanly on cancellation.

Plugin wire compatibility

Backward-incompatible plugin protocol changes usually require updating the plugin API version and checking compatibility behavior for older plugins.

CBOR byte decoding

CBOR byte payloads may decode into different Go types depending on decoder behavior. Avoid brittle direct assertions when helper functions exist to normalize input.

Parsed-but-unused config fields

Config fields should not be added unless their behavior is implemented. Otherwise users can set them successfully and still see no effect.

Generated commands and OpenAPI behavior

Spec-loading or command-generation changes can break existing workflows indirectly. Check for backward compatibility in generated command shape, naming, argument expectations, and operation discovery.

Spec and operation cache invalidation

A recurring source of fixes: stale generated commands after config or spec changes. Caches are keyed by base URL, operation base, server variables, and raw spec hash. Changes to api edit/api connect, profiles, base URLs, or spec loading must invalidate or refresh the relevant caches, and concurrent api connect runs must not clobber each other's entries.

OAuth and auth concurrency

Token refreshes must stay serialized; discovery transports and OAuth state are shared across generated commands. Changes to internal/auth should be checked for refresh races, cache-key reuse across profiles/issuers, and browser-launch behavior ($BROWSER handling, Linux flows).

v1 → v2 migration of persisted state

Changing config schema, cache key formats, or credential storage breaks existing user state unless migrated. Past fixes migrated legacy token cache keys, legacy tls.cert/tls.key fields, and added migration warnings. New formats need a mapping from the old one or an explicit, tested migration error.

Command surface and help contracts

internal/cli/command_surface_test.go locks the built-in command tree, help grouping, error-message UX, and removed pre-release names. Adding, renaming, or regrouping commands should update it deliberately — treat unexplained churn there as a regression signal.

Output formatting

Intentional formatter changes should usually come with targeted regression coverage. Unintentional output drift is a common source of user-visible regressions.

Auth, pagination, filtering, and caching flows

Changes in these areas often regress behavior only in realistic end-to-end paths. Review interactions, not just isolated helpers.

For auth, redaction, and cache metadata changes, check both positive behavior and negative leakage boundaries: where credentials are applied, where they must not be applied, what gets persisted, and what appears in errors or traces. Keep findings tied to the PR's changed paths.

internal/secrets holds the central allow-lists for recognizing credential-like fields; new secret-ish field names belong there, not in ad-hoc checks. Past fixes covered suffixed fields (confirm_password), URL userinfo and query credentials in network errors, and avoiding false positives on ordinary fields like max_tokens.

Test buffer races

Tests that share a bytes.Buffer across concurrent writers can hide data races, especially when subprocess stderr/stdout is wired into test buffers.

Verification Hints

• Run the AGENTS.md verification ladder: narrowest touched package first, then go test ./..., then -tags=integration for CLI/plugin behavior changes, and -race when concurrency, subprocess handling, or shared buffers are touched.
• Run go run ./cmd/restish-docgen --check when CLI surface or help text changed; CI fails on stale generated docs regions.

Example Findings

Good examples illustrate the shape of a finding without hard-coding one exact response style:

• "This changes plugin message semantics but does not bump the plugin API version, so older plugins may silently misbehave."
• "The new timeout returns early, but the goroutine reading from the pipe still blocks forever if the subprocess never responds."
• "The happy-path test passes, but there is no coverage for malformed config or partial server responses."
• "This adds a user-facing flag, but the docs site does not appear to mention it."